Key Takeaways
- For most multi-step workflows that call tools and produce structured output, GPT-6.1 Terra at high effort and Claude Sonnet 5.5 at medium effort give the best balance of cost and reliability in late October 2026.
- In a nine-task DevOps test, GPT-6.1 Terra (high) passed all nine tasks for $0.0626, Sonnet 5.5 (medium) passed all nine for $0.0665, and Claude Opus 5.5 passed all nine for $0.2732.
- For high-volume classification, routing and data extraction, Claude Haiku 5.5 and GPT-6 Luna are the cheapest capable options at $0.10 per million input tokens and $0.50 per million output tokens.
- For long, complex coding work, GPT-6.1 Sol offers strong value: about $0.72 per benchmark task on Artificial Analysis, against $3.26 for GPT-6 Astra.
- The cheapest model per token is not always the cheapest per task. GPT-6 Luna cost under half a cent for the test but failed three of nine tasks.
- A router setup, where a cheap model handles easy steps and escalates hard ones, usually beats picking a single model.
Using Agentic AI workflows in business settings. Image source: Alius Noreika / AI
The best cost-efficient AI model for workflow automation in late 2026 depends on what each step of the workflow does. For general multi-step jobs that call tools, edit files or produce structured output, GPT-6.1 Terra at high effort and Claude Sonnet 5.5 at medium effort deliver near-flagship reliability for a fraction of flagship prices. For simple, high-volume steps such as sorting tickets, tagging documents or pulling fields out of emails, Claude Haiku 5.5 and GPT-6 Luna cost about a tenth as much.
Price per token is only the starting point. Reasoning tokens, failed attempts and retries all add to the bill, so the model that finishes a task correctly on the first pass is often the cheapest in practice. The tables below combine list prices with hands-on test results to show where each model offers real value.
AI Model prices for workflow automation
| Model | Input per 1M tokens | Output per 1M tokens | Cached input | Best suited to |
|---|---|---|---|---|
| Claude Haiku 5.5 | $0.10 (prompts up to 100K tokens); $0.50 above | $0.50; $2.50 above 100K | $0.01 | Classification, routing, extraction, subagents |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 | Simple, high-volume steps |
| Gemini 3.8 Flash | $0.75 | $3.75 | — | Fast multimodal tasks |
| GPT-6.1 Terra | $1.00 | $5.00 | $0.05 | General tool-using workflows |
| Meta Muse Spark 1.3 | $1.25 | $4.25 | $0.15 | General tasks, agent apps |
| Claude Sonnet 5.5 | $2.00 | $10.00 | $0.20 (reported cut to $0.10) | Reliable multi-step workflows, coding |
| GPT-6.1 Sol | $2.00 | $10.00 | $0.10 | Long coding and analysis |
| Gemini 4 Argon | $2.00 until Nov 30; $4.00 after | $10.00; $20.00 after | $0.10; $0.20 after | Very long documents |
| Claude Opus 5.5 | $4.00 | $20.00 | $0.20 | Hardest tasks needing first-pass accuracy |
| GPT-6 Astra | $10.00 | $50.00 | — | Frontier tasks where cost matters less |
A few pricing rules change the math. Reasoning tokens are billed as output on every platform listed here. OpenAI charges double for input and 1.5 times for output on prompts above 272,000 tokens.
Anthropic and OpenAI both offer batch processing at half price, and OpenAI’s Flex tier carries the same discount. Gemini 3.8 Flash is set to double in price on January 1, 2027, to $1.50 input and $7.50 output. Reports say Anthropic cut Sonnet 5.5’s cache read price to $0.10 on October 7, though its documentation still listed $0.20 at the time of writing, so check before planning around the lower figure.
Real test results: cost per workflow run
Computing for Geeks ran the same nine-call DevOps workflow on a range of models: writing Kubernetes manifests, Bash scripts and OpenTofu infrastructure files, then checking whether the results actually worked. The table shows how many tasks each model completed correctly and the total cost of the run.
| Model and setting | Tasks passed (of 9) | Total cost |
|---|---|---|
| GPT-6.1 Terra (high) | 9 | $0.0626 |
| GPT-6.1 Sol (medium) | 9 | $0.061 to $0.064 |
| Claude Sonnet 5.5 (medium) | 9 | $0.0665 |
| GPT-6 Sol | 9 | $0.0841 |
| Claude Sonnet 5.5 (high) | 9 | $0.1230 |
| Claude Opus 5.5 | 9 | $0.2732 |
| GPT-6.1 Terra (medium) | 8 | $0.0429 |
| GPT-6.1 Terra (no reasoning) | 5 | $0.0170 |
| GPT-6 Luna | 6 | $0.0036 |
Two findings stand out. First, dropping GPT-6.1 Terra from high to medium effort saved only about 30%, not the 50% one might expect, because reasoning-token use did not shrink as much as the lower setting implies, and the medium run still missed one formatting check.
Second, GPT-6 Luna was almost free but failed all three Kubernetes tasks, so any workflow relying on it would need retries or human fixes that erase the savings. Claude Opus 5.5 got everything right, but at more than four times the cost of Terra or Sonnet for the same result.
Cost per task on broader benchmarks
Artificial Analysis measures what each model spends to finish its full set of benchmark tasks, which captures token-hungry reasoning better than list prices.
| Model | Approximate cost per index task | Intelligence Index |
|---|---|---|
| GPT-6 Luna (max) | $0.07 | — |
| GPT-6.1 Terra | $0.40 | 48 |
| GPT-6.1 Sol | $0.72 | 52 |
| GPT-6 Sol | $1.06 | — |
| Claude Opus 5.5 (high) | About $1.82 | 54 |
| Gemini 4 Argon (intro price) | $1.99 | 53 |
| GPT-6 Astra | $3.26 | 53 |
GPT-6.1 Sol stands out for coding value. At high effort it scored 75.2% on DeepSWE for about $0.65 per task, while GPT-6 Astra scored 74.1% at around $4.43 per task. Nearly the same result for roughly a seventh of the price makes Sol hard to beat for long software tasks.
Claude Haiku 5.5: the new budget option
Anthropic released Claude Haiku 5.5 on October 7, 2026, and built it specifically for classification, routing, extraction and work as a subagent inside larger systems. It costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, rising to $0.50 and $2.50 above that. Cache reads cost $0.01, and batch processing halves the price again.
Anthropic says Haiku 5.5 is about 75% cheaper than Haiku 4.5 on average and 90% cheaper for prompts under 100,000 tokens, which covered about 90% of Haiku 4.5 requests. One catch: its new tokenizer produces roughly 30% more tokens for the same text than Haiku 4.5, which trims some of the savings.
Haiku 5.5 supports a one-million-token context window and 128,000 output tokens, and Anthropic reports a 72.4% score on OSWorld 2.1 for computer use, far ahead of GPT-6 Luna’s 48.9%. That figure comes from Anthropic, not independent testing.
Readers who followed earlier budget models, such as GPT-5.6 Luna, will notice how far prices have fallen in a single generation. Our previous budget model comparison shows the starting point.
Open and lower-cost alternatives
Several models outside the big three labs also compete on price. Zhipu’s GLM-5.3 lists at $1.40 input and $4.40 output per million tokens. Mistral Large 4 was on sale at $0.68 and $2.09, half its $1.36 and $4.18 list price. Moonshot’s Kimi K3 costs $3 and $15. Meta’s Muse Spark 1.3 drops to $0.10 input and $0.20 output in its Contributor mode, where Meta may train on the traffic. That is a trade-off many businesses will not accept for sensitive data. For background on one of these challengers, see our deep dive into Moonshot AI and Kimi.
Which model to choose for each workflow type
| Workflow step | Best value pick | Why |
|---|---|---|
| Ticket routing, tagging, sentiment, field extraction | Claude Haiku 5.5 or GPT-6 Luna | Lowest price; simple tasks rarely need deep reasoning |
| Multi-step tool use, report generation, data cleanup | GPT-6.1 Terra (high) or Claude Sonnet 5.5 (medium) | Full pass rate in testing at a low cost per run |
| Long coding tasks and refactors | GPT-6.1 Sol (high) or Claude Opus 5.5 (medium) | Strong coding scores; Sol is cheaper, Opus more consistent |
| Very long documents or archives | Gemini 4 Argon, ideally before December 1 | One price across a 1M-token window |
| Overnight batch processing | Any of the above via batch or Flex tiers | Half price when results can wait |
How to cut workflow costs further
The most effective setup uses more than one model. A cheap model such as Haiku 5.5 handles the first pass and decides whether a task needs escalation, while a mid-tier model like Terra or Sonnet handles the harder steps, and a flagship like Opus 5.5 only sees the small share of tasks that fail lower down.
Validating each output with a schema check, a test or a simple rule catches failures before they spread through the pipeline, which matters most when using the cheapest models.
Prompt caching is the other major lever. Keeping fixed instructions, examples and reference material at the start of every prompt lets cache discounts of 90% or more apply to most input tokens. Setting effort levels per step, rather than running every call at maximum, keeps reasoning tokens in check.
Our guide to AI in business workflows covers where automation pays off first, and our test of coding cost savings across models shows how quickly per-task differences add up at scale.
If you are interested in this topic, we suggest you check our articles:
- How Affordable Is ChatGPT 5.6 Luna?
- GPT-5.6 Luna vs Gemini 3.6 Flash vs Claude Opus 5
- How AI Is Transforming Business Workflows
- Grok 4.5 vs Claude Opus: Coding Cost Savings
- Moonshot AI Company Deep Dive: Kimi
Sources: Computing for Geeks – GPT-6 model tests, Computing for Geeks – Claude Opus 5.5, Anthropic – Claude Haiku 5.5 documentation, gori.me, eesel AI
Written by Alius Noreika
